Using Support Vector Machine approach for

forecasting the failure of the Tunisian companies

Yousri HENCHIRI and Salwa BENAMMOU and Taher HAMZA

Abstract In this study we look to analyze, the financial determinants of failure and the
ability of predictive models to anticipate the occurrence of this risk. This gait grants a
preventive character through the establishment of precocious alert system rather than that
of making-decision. The background in this field suggested various conceptions tools for
the prediction of failure, such as discriminate analysis, logistic regression, artificial neural
network, genetic algorithms... As part of this research, we propose to test the predictive capacity of the SVM model concerning distressed companies. This model, that is a learning
algorithm class, initially conceived for discrimination, has been applied on two samples
of small and medium Tunisian companies: a training sample and a test sample. The gridsearch technique, using the cross-validation, is used to find out the best parameter value
of SVMs kernel function. Finally, we empirically show that the SVM model gives high
accuracy rate than other approaches being tested in this domain.
Key words: Failure, Prediction model, financial determinants, Support Vector Machine,
Grid-search

1 Introduction
The analysis of the determinants of failure shows that these latters are of various natures:
macroeconomic, strategic and organizational and finally financial. Besides, we noticed that
most studies adopt a multi-referential methodology enrolling in a double optical: the one
Publi
e in the Book of Short Papers, 7th Scientific Meeting of CLAssification and Data Analysis Group
(CLADAG) of the Italian Statistical Society, Catania 2009.

Yousri HENCHIRI and Taher HAMZA

Institut Superieur de gestion, Universite de Sousse, Tunisia., e-mail: [email protected] and e-mail:
[email protected]
Salwa BENAMMOU
Faculte de Droit & des Sciences Economiques & Politiques, Universite de Sousse, Tunisia., e-mail: [email protected]

Yousri HENCHIRI and Salwa BENAMMOU and Taher HAMZA

is that of optimization forecasting financial distress approach and the other is that of conceptualization. In this study we look to analyze, the financial determinants of failure and
the ability of predictive models to anticipate the occurrence of this risk. The background
in this field suggested various conceptions tools for the prediction of failure, such as multiple discriminate analysis (MDA) [3]-[5], logistic regression (LOGIT) [9]-[5], artificial
neural network (ANN) [3]-[5], decision tree [4]-[11], genetic algorithms [11]-[7]. Recently, a learning algorithm class, initially conceived for discrimination under the name
of, Support Vector Machine (SVM) has been also introduced [8]-[6]-[10]. We look to apply this forecasting approach of the risk of the financial distress and provide a new tool
improving its prediction accuracy. Developed by Vapnik (1998), SVM method is gaining
popularity thanks to many attractive features and excellent generalization performance on
a wide range of problems. In addition, since the optimal parameters search of SVM plays
a crucial role in building a distress prediction model with high accuracy and stability; we
apply 5-fold cross validation and a grid-search technique in order to identify the correct
value parameter of the kernel function of SVM. We use LIBSVM software [2] to conduct
SVM experiment.

2 Research methodology and analysis results

TThe database used in this study, is composed of 120 firms [ see [5], were the sample
is composed of two groups of firms : The first is the small and medium Tunisian companies (SMTC) such in a state of failure and the second group is sane SMTC]. The research
methodology goes around 5 phases: a selection of discriminatory 28 financial ratios (inspired by previous researches). over a period of two successive years, a test of Equality
of Group means ratios between the two companies populations (sane and distressed), developing a matrix of correlations between ratios, using the principal component analysis
(PCA) to reduce the size of each feature and to sequentially optimize a choice more limited
of ratios and finally a SVM process is applied in an optical forecasting financial distress
to 1 and 2 years of the event.following this process, we did a set of experiments with different kernel functions such as the linear, RBF, polynomial, and sigmod in order to see
the quality of generalization for each kernel function. We apply a grid-search to find the
best pair (C, gamma) for each kernel function using 5-fold cross-validation. In order to increase efficiency, we try exponentially growing sequences of (C, gamma) to identify good
parameters (C = 25 , 23 , ..., 215 ; gamma = 215 , 212 , ..., 212 ).
After the optimal (C, gamma) is found, the whole training data is trained using the SVMs
with different kernels and the best parameters to generate the final models.
Table 1 compares prediction performance of the SVM models using four different kernel
functions. We note, in case of the RBF kernel, the prediction accuracy of tested data is
turned out to be 95%, while that of the training data is 100%. As shown in Table 1, in
terms of recognition accuracy, the RBF kernel gets the best result (100%) followed by the
polynomial kernel (100% when d = 2), the linear kernel (97.5%) and finally the Sigmod
kernel (80%). The highest prediction accuracy (97.5%) obtained by the linear kernel and
the polynomial kernel followed by the RBF kernel (95%) and the sigmod kernel (92.5%).
Table 1 also shows that the underfitting or the overfitting problem has occurred in case of
the polynomial kernel function. This is a typical overfitting phenomenon in machine learn-

Using Support Vector Machine approach for forecasting the failure of the Tunisian companies

Table 1 Performance of SVM Kernel on each optimal (C, ).

Kernel Function

Gamma

d
N/A
N/A
1
2
3
4
N/A

Linear
RBF
Polynomiala

25
29
211

26
29
212

Sigmodea

21

26

Prediction accuracy (%)

Training data
Test data
97.5% (78/80)
100% (80/80)
98.75% (79/80)
100% (80/80)
100% (80/80)
100% (80/80)
80% (64/80)

97.5% (39/40)
95% (38/40)
97.5% (39/40)
97.5% (39/40)
97.5% (39/40)
95% (38/40)
92.5% (37/40)

Parameter r in set to 0.

ing. Therefore, we need to make extra efforts to find the best value of the degree d in the
polynomial kernel SVM model. As shown in Table 1, the RBF kernel achieves the highest
prediction accuracy better than that in other kernels. In addition, overfitting is unlikely to
occur with the RBF kernel function.
In the same perspective followed in the analysis of the data 1-year before the failure that
we have processed the data 2-year the failure. Table 2 compares prediction performance of
the SVM models using four different kernel functions. In case of the RBF kernel, the prediction accuracy of tested data is turned out to be, 92.5% while that of the training data is
98.75%. As shown in Table 2, the RBF kernel gets the best recognition accuracy (98.75%)
followed by the polynomial kernel (98.75% when d = 2), the linear kernel (97.5%) and
finally the Sigmod kernel (95%). Also, the linear kernel and the Sigmod kernel obtained
the best prediction accuracy of tested data (97.5%), followed by the polynomial kernel
(95%) and the RBF kernel (92.5%). The obtained results are relatively satisfactory for a
Table 2 Performance of SVM Kernel on each optimal (C, ).
Kernel Function

Gamma

Linear
RBF
Polynomiala

23
211
211

212
212
212

Sigmodea

27

26

N/A
N/A
1
2
3
4
N/A

Prediction accuracy (%)

Training data
Test data
97.5% (78/80)
98.75% (79/80)
98.75% (79/80)
100% (80/80)
100% (80/80)
100% (80/80)
95% (76/80)

97.5% (39/40)
92.5% (37/40)
95% (38/40)
90% (36/40)
90 % (36/40)
90% (36/40)
97.5% (39/40)

Parameter r in set to 0.

near horizon (1-year) for which recognition accuracy varies respectively from 80% and
100% according to SVM model applied respectively on a training sample. Also, the forecasting accuracy varies respectively from 92.5% and 97.5% according to SVM model
applied respectively on a test sample. These results are better via the rates that are found
with MDA and LOGIT model. On the basis of the same sample, we compare the results of
the SVM model of our study with those obtained by the ANN, MDA and LOGIT models

Yousri HENCHIRI and Salwa BENAMMOU and Taher HAMZA

from the study of [5]. Table 3 also summarizes the prediction performance of SVM approach, ANN, MDA and LOGIT model. As shown in Table 3, SVM and ANN are slightly
outperforms LOGIT model and MDA for tested data.
Table 3 The best prediction accuracy of SVM, ANN, LOGIT model and MDA (%).
Training data
Data (T-1) Data (T-2) Average
SVM
ANN
LOGIT
MDA

100%
95%
93.75%
93.75%

98,5%
100%
95%
96.75%

99.25%
97.5%
93.175%
95%

Test data
Data (T-1) Data (T-2)
95%
95%
92,5%
72.5%

92,5%
92.5%
80%
62.5%

Average
93.75%
93.75%
86.25%
66.875%

Table 3, compares the best performance of SVM, ANN, MDA and LOGIT in the training
data and tested data and show a superiority of the SVM method in relation to ANN and
to the traditional approaches parametric. In relation to the remoteness of the forecasting
horizon, all models tested in our study experienced know, an average, improvement in the
phase and deterioration in their external validation on a test sample. We can conclude, nevertheless, that the discrimination algorithms are a good preacher of the risk of distress to a
close horizon to the failure, especially if the forecasting is made on the horizon of 1-year.
Another particularity of our results is the performance of the SVM for kernel RBF rather
than other functions kernel to a close horizon to the failure.

3 Conclusion
The application of this technique to two samples of firms (sane and distressed) permitted us
to obtain significant results and to propose a forecasting model more appropriate in terms
of quality of generalization. The explanation for these results is related to causes rather
than economic statistics since the SVM modeling does not require restrictive assumptions
than those for traditional approaches. Concerning the results we noticed a significant difference from firms distressed and sane in terms of the business activity, the needs cleared
by their exploitation, in debts rate, profitability, financial imbalance, liquidity and solvency.
These reports are consistent with 1 and 2 years before the failure. Finally, the anticipation
of the risk of financial distress of firms with SVM enabled us to achieve significant results
and propose a suitable prediction model with a predictive capacity between 92.5% and
97.5% a 1-year before the failure.

References
1. Vapnik, V.N.: Statistical learning theory. New York: Springer (1998)
2. Hsu, C.-W., Chang, C.-C., Lin, C.-J.: A practical guide to support vector classification. Tech. rep.,
Department of Computer Science and Information Engineering, National Taiwan University(2008)

Using Support Vector Machine approach for forecasting the failure of the Tunisian companies

3. Altman, E.I., Marco, G., Varetto, F.: Corporate distress diagnosis: comparisons using linear discriminant analysis and neural networks (the Italian experience). Journal of Banking and Finance. Appl.
18(3), 505529 (1994)
4. Frydman, H.E., Altman, E.I., Kao, D.: Introducing recursive partitioning for financial classification:
The case of financial distress. The Journal of Finance. Appl. 40(1), 269291 (1985)
5. Hamza, T., Baghdadi, K.: Profil et determinants financiere de la defaillance des P.M.E Tunisiennes
(1999-2003). Banque et Marches. Appl. 93(Mars-Avril), 4562 (2008)
6. Hua, Z.-S., Wang, Y., Xu, X.-Y., al.: Predicting corporate financial distress based on integration of
support vector machine and logistic regression. Expert Systems. Appl. 33, 434440 (2007)
7. Liang, L., Wu, D.-S.: An application of pattern recognition on scoring Chinese corporations financial
conditions based on back propagation neural network. Computers & Operations Research. Appl.
32(5), 11151129 (2005)
8. Min, J.-H., Lee, Y.-C.: Bankruptcy prediction using support vector machine with optimal choice of
kernel function parameters. Expert Systems. Appl. 28, 603614 (2005)
9. Mossman, C.E., Bell, G.G., Swartz, L.M., Turtle, H.: An empirical comparison of bankruptcy model.
Financial Review. Appl. 33, 3554 (1998)
10. Sun, J., Lin, H.: Data mining method for listed companies financial distress prediction. Journal of
Knowledge-Based systems. Appl. 21, 15 (2008)
11. Varetto, F.: Genetic Algorithms applications in the analysis of insolvency risk. Journal of Banking
and Finance. Appl. 22(10-11), 14211439 (1998)